Home / Curated / Algorithms / Equalized-odds postprocessing for binary predictors

Equalized-odds postprocessing for binary predictors

Algorithms AIMech:EqualizedOddsBinaryPostprocessing DRAFT

Aliases: Hardt equalized-odds binary postprocessing

A randomized postprocessing algorithm that uses a binary predictor and protected-group membership to minimize expected label loss subject to equal false-positive and true-positive rates across groups, within predictors derived from those inputs.

Identity Notes & Concept Scope This record selects the Section 4.1 binary-input linear program. It is distinct from the EqualizedOdds criterion and from the continuous-score threshold construction in Section 4.2. Equal-opportunity-only constraints define a relaxation, not an alias here. All 442 pinned AIO labels/scoped synonyms and 99 preceding native identities were compared, together with hidden/ignored-inclusive searches of both checkouts; source citations and pending frontier mentions were distinguished from existing concept identities.

Claims & Grounded Evidence

1 Claim

Mechanism #method-scope

For binary groups and binary predictions, optimize four probabilities of outputting one, indexed by the original prediction and group. The joint distribution of original prediction, group and target supplies linear loss and parity constraints. The feasible group error-rate pairs form the convex hull of the original classifier, its complement and the two constant classifiers; their intersection supplies equalized-odds solutions. Prediction uses the original output and group, with randomization as needed.

supports Source: Equality of Opportunity in Supervised Learning · Locator: Section 4, Definition 4.1 and Section 4.1, equations (4.1)–(4.3), Proposition 4.4, PDF pp. 5–7; Section 6 and Discussion for distributional limitations

⚠️ Curated Limitations & Negative Results

Optimality is over randomized predictors derived from the supplied binary prediction and group for the specified distribution and loss, not every possible predictor using raw features.
Fitting needs target labels and group information; deployment needs group membership. Empirical estimates and distribution changes can invalidate population parity.
The four-parameter statement assumes two groups and binary predictions; it does not assert that a general score-valued postprocessor has four parameters.
Equal conditional error rates do not establish equal accuracy, individual fairness or a universal fairness judgment. The printed proof is not reproduced as implementation code.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:BinaryClassification` Binary Classification	related_match	Postprocesses binary classification decisions; the algorithm and the task are related at different abstraction levels, so this is not an equivalence or subclass assertion.

Evidence Sources & Primary Literature

1 Source

Source ID	Publication / Document	Type / Access	Reference & Checksum
`original-paper`	Equality of Opportunity in Supervised Learning Moritz Hardt, Eric Price, Nathan Srebro	primary paper full_text	ID: arXiv:1610.02413 sha256:e92b4f579180…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-11T04:57:48Z

Updated: 2026-10-11T04:57:48Z

Source File: data/curated/EqualizedOddsBinaryPostprocessing.yaml